Federal Reserve lacked access to Anthropic Mythos for 3+ months despite cybersecurity warning to banks

In April 2026, the Federal Reserve and US Treasury convened an extraordinary meeting with major bank CEOs to warn them that Anthropic's Claude Mythos Preview posed an unprecedented cybersecurity threat due to its exceptional ability to identify software vulnerabilities. The Fed itself, however, did not have access to the model and remained locked out for at least three months—through mid-July—even as Anthropic expanded Project Glasswing access from roughly 50 initial organizations to over 150 across 15 countries. Federal Reserve Chair Kevin Warsh testified to Congress on July 15 that the Fed was still seeking access to Mythos and other advanced AI models.

Anthropic had granted early access to roughly 50 organizations including JPMorgan Chase, Amazon, Apple, and Google as of April, but the Federal Reserve—arguably the most systemically important financial institution in the world—was not among them. In June, Anthropic temporarily disabled Mythos 5 and its broader Fable 5 variant due to export control directives citing national security, though Commerce Secretary Howard Lutnick later granted selective access to trusted partners. The full export restrictions were eventually lifted.

For architects and policymakers, this incident exposes a critical gap: the central bank tasked with protecting US financial stability cannot quickly access the security assessment tools it needs, while the vulnerability identification capability spreads widely. The three-month access lag raises questions about how vulnerability disclosure and remediation workflows should be coordinated when frontier AI models become defensive cybersecurity assets.